Agentic AI musings 

I’ve been talking about these shifts in the Mercer HR Tech Forum (along with my co-hosts Harry West and Tara Cooper), but I wanted to share a more accessible take here. 

Consider this a field note on where Agentic AI in HR is headed, what’s changed since early predictions, and why HR leaders need to think differently about managing not just people, but intelligent digital partners.

In January 2024, I published an article called Generative AI and large language models drive the creation of human-agent teaming on the Mercer Talent All Access Portal (TAAP+). Back then, I used the term “digital co-workers” — today, I tend to say “human-agent teaming,” which feels closer to the reality of how people and AI agents work together.

In that earlier piece, my example was a ServiceNow assistant “talking” to a Workday assistant to retrieve HR information or complete a task on behalf of an employee.

At the time, my thinking was that it didn’t make sense for organizations to have one “uber” chatbot that could do everything. Instead, each vendor would have its own primary interface — the one most employees interact with — plus other assistants with deep expertise in specific domains or systems — in my example, ServiceNow as the front door, and Workday with the HR knowledge and permissions to act.

That still makes sense. What’s changed is that even within a single product, one assistant isn’t enough. To get the accuracy and depth required for certain tasks, vendors are creating multiple specialized agents — each built for a narrow set of skills. We’ve seen a flood of announcements in the past six months as this model takes hold.

The more specialized agents you have, the more important orchestration becomes — not just within a platform, but across platforms. In the ServiceNow or Workday examples, their chatbots/digital assistants can orchestrate their own agents. That’s progress. However, from an enterprise perspective, those agents still need to work with agents in other systems. Protocols have emerged like Model Context Protocol (MCP) from Anthropic and Agent2Agent (A2A) from Google to provide a more standardized way to orchestrate workflows and tasks across vendor products. Both ServiceNow and Workday, for example, have both indicated that they would support A2A.

For HR, this means orchestration protocols will become critical for ensuring agents across payroll, benefits, recruiting, and other platforms can actually work together.

Another major change since I first wrote about this space is the reasoning capability of the latest large language models. Instead of being limited to what the LLM has been pre-trained to know or do, it can now create its own plan for how to fulfill a request.  This means an LLM does not have to be pre-trained for every use case you might want them to address — the model can figure out a path forward, often explaining its steps along the way.

Because of the orchestration protocols, they can also use different tools to accomplish tasks.  This is quite a leap forward in usefulness. But there is a catch. There are limits to how much one agent can orchestrate other agents to do, at least today. 

This video from Patrick Dougherty is worth watching and offers a better understanding of the challenges of building Agentic AI systems.

One of the conclusions from Patrick’s story on building these Agentic AI systems is that an Orchestrator (Manager) agent can only reasonably leverage 5-8 other agents to do work towards a particular outcome. Beyond that, coordination breaks down — just like a human manager with too many direct reports.

For those of us versed in the HR perspective, this may feel familiar.  Just like human managers, agent “managers” need a manageable span of control and the right resources to succeed.  This post from Anthropic Engineering tells a similar story.

That’s why the concept of an “Agent System of Record” caught my eye when Workday introduced it. At first, I was skeptical, as it felt like a marketing play to lean into AI hype. But the more I’ve watched Agentic AI in HR evolve, the more I think they may not have gone far enough.

If you believe the more aggressive predictions, enterprises could someday have thousands or even millions of agents.  There has been some press on the notion of agent swarms , for instance.  Based on AI technology, as it stands today, that would be pure chaos for most enterprises.  The most effective pattern I’ve seen — and one that mirrors human organizations — is to group agents into teams with a clear leader (orchestrator) and goals.

Teams will still need to coordinate with each other.  What could also make sense is a hierarchical agent org chart (possibly mixed with a human worker org chart depending on the work) that has the right span of control for the work that AI agents perform. 

Those goals could look a lot like the ones we assign to human teams in a performance management system. If an agent team isn’t meeting its targets, you may need to retrain one or more agents, swap in different capabilities, or redesign the workflow.

This is where HR’s expertise becomes critical. The principles we already use for selecting talent, assigning work, developing skills, and measuring performance apply just as well to 

Right now, most enterprise agent management is focused on plumbing: how to connect systems, enable workflows, and make sure data moves where it needs to go. That is necessary, but not sufficient. 

To get the most out of human-agent teaming, we need to think more strategically about how we design and manage hybrid digital workforces. This includes how we select, assign work to, organize, develop, and manage agents in ways like, and alongside, our human workforce (while still accounting for the differences in our agent workforce). 

HR is uniquely positioned to lead here, because this is what we already do with the human workforce. The technology is advancing quickly. The management mindset needs to catch up.

In conclusion, as Agentic AI continues to evolve at a rapid pace, the role of HR in shaping the future of work becomes more critical than ever. The shift from viewing AI as mere tools to recognizing them as intelligent digital partners demands a new management paradigm—one that blends traditional HR expertise with innovative strategies for orchestrating and developing hybrid human-agent teams. By applying proven principles of talent management, performance measurement, and organizational design to these emerging digital workforces, HR leaders have a unique opportunity to drive meaningful impact and ensure that both human and agent collaborators thrive together. Embracing this challenge will not only enhance operational efficiency but also position HR at the forefront of the next frontier of transformation.